Image generation models on Mac: Draw Things, DiffusionBee and ComfyUI

The Draw Things models folder on a Mac is ~/Library/Containers/com.liuliu.draw-things/Data/Documents/Models, inside the app's own sandbox container. Every model you download or import in Draw Things lands there, with the LoRAs, text encoders and other parts the models need, and it stays until you delete it in the app.

DiffusionBee keeps its models in a hidden folder in your home folder, ~/.diffusionbee. ComfyUI keeps them in its models folder: for Comfy Desktop, that's the shared library in ~/ComfyUI-Shared, used by every installation you create.

These folders get big fast. One FLUX image model with the files it needs is 16 to 34 GB, depending on the app and the precision, and the largest image models pass 50 GB. This post shows what's in each folder, what each file is for, and how to remove models without losing your own work.

Where image generation apps keep their models

Each app keeps its own models in its own format, so the same model downloaded in two apps takes its space twice. Here's where each one puts them.

App	Where its models live	What else is nearby	Remove models in
Draw Things	~/Library/Containers/com.liuliu.draw-things/Data/Documents/Models	The app's other data, in the same container	Settings, Model, Manage
DiffusionBee	~/.diffusionbee, in downloaded_assets and imported_models	Your generated images, in images	Finder, with the app quit
Comfy Desktop	~/ComfyUI-Shared	Installations, in ~/ComfyUI-Installs	Finder, or the app's Uninstall for an installation
ComfyUI from GitHub	The models folder of wherever you installed it	Your outputs and custom nodes	Finder
Diffusers and other Python tools	~/.cache/huggingface/hub	Every other Hugging Face download	Finder, or Hugging Face's own tool

Why is Draw Things' folder inside Library/Containers?

Draw Things is a sandboxed app, so macOS gives it a container of its own and keeps its files there, away from other apps. Its Documents folder is the container's, not yours. Containers and Group Containers on a Mac explains how that works and why Finder shows these folders by app name.

Draw Things' own command-line tool uses the same Models folder by default, so a model you download once serves both.

How to open these folders in Finder

In Finder, choose Go › Go to Folder, type the path, such as ~/Library/Containers/com.liuliu.draw-things/Data/Documents/Models, and press Return. The tilde stands for your home folder, and folders whose names start with a period, like .diffusionbee, are hidden but open this way.

To see what a folder takes, select it and choose File › Get Info. In list view, View › Show View Options has a Calculate all sizes option that shows every folder's size, so you can sort by it.

What Storage settings shows

Choose Apple menu › System Settings, click General, then Storage. There's no category for AI models. Apple describes the Documents category as files in your home folder that aren't in other categories, with buttons for large files, so a large model file in a visible folder can appear there.

Models in hidden folders and app containers are easier to find in Finder. Storage settings works the same way in macOS 26 Tahoe and macOS 27 Golden Gate.

The Draw Things models folder, file by file

Open the Draw Things models folder and you'll see .ckpt files with long names. That's Draw Things' own format: every model, whether downloaded or imported, is stored as one or more of these files.

What do the file names mean?

The end of a file's name says how precisely the model is stored, which decides most of its size. Draw Things' model menu marks the same versions with the words in the second column.

Ending	Shown in Draw Things as	What it is
_f16.ckpt	Exact, or no mark for smaller models	Full 16-bit weights, the largest
_q8p.ckpt	No mark for many large models, or 8-bit	8-bit, the default for many large models
_q6p.ckpt	6-bit	Smaller, with a little loss
_q5p.ckpt	5-bit	Smaller again
_svd.ckpt at the end	SVDQuant	A version quantized with a method called SVDQuant
_lora_f16.ckpt	A LoRA	An add-on that changes a model's style or subject

For FLUX.1 [schnell], for example, Draw Things lists the 8-bit flux_1_schnell_q8p.ckpt as the plain model, at 12.3 GB, and the 5-bit flux_1_schnell_q5p.ckpt at 9.4 GB.

Why is one model several files?

A modern image model is several networks working together: the main model, a text encoder that reads your prompt, and an autoencoder that turns the result into pixels. Draw Things stores each as its own file.

FLUX.1 [schnell] in its default 8-bit version is four files: the 12.3 GB model, a 3.9 GB T5 text encoder, a 0.25 GB CLIP text encoder and a 0.17 GB autoencoder. Together that's about 16.6 GB.

Models share their parts

Models from one family use the same text encoders and autoencoder. FLUX.1 [dev] uses the same three helper files as [schnell], so adding it costs only its own 12.3 GB model file. SDXL and FLUX share the CLIP encoder too.

That's why deleting a model may free less than its listed size, and why removing files by hand is risky: a text encoder you remove may be one another model still needs.

The large helper files

Most of the big files in the folder are main models, named after them, such as sd_xl_base_1.0_f16.ckpt. The helper files are named after what they are instead, and a few are large in their own right:

File	What it is	Used by	Size
clip_vit_l14_f16.ckpt	CLIP text encoder	SDXL and FLUX.1 models	0.25 GB
open_clip_vit_bigg14_f16.ckpt	Larger CLIP text encoder	SDXL models	1.39 GB
t5_xxl_encoder_q6p.ckpt	T5 text encoder, 6-bit	FLUX.1 models	3.88 GB
flux_1_vae_f16.ckpt	Autoencoder	FLUX.1 and Z Image	0.17 GB
umt5_xxl_encoder_q8p.ckpt	UMT5 text encoder, 8-bit	Wan video models	6.88 GB
qwen_2.5_vl_7b_f16.ckpt	Qwen text encoder, 16-bit	Qwen Image, Exact versions	14.16 GB

A helper file stays useful as long as any model of its family remains. Once you've deleted every Wan model, for example, the 6.9 GB UMT5 encoder has nothing left to serve.

Imported models and LoRAs

When you import a model, Draw Things converts it into its own format and saves the result in its models folder. The file you imported stays where it was, so until you remove it, the model takes its space twice.

Imported Stable Diffusion models also bring their own copy of the text encoder, named after the model, such as dreamshaper_v8_clip_vit_l14_f16.ckpt. LoRAs you import or train are stored in the same folder.

How to delete models in Draw Things

Draw Things removes models in its own window, which knows which files each model uses:

1. In Draw Things, open Settings.
2. Click the Model menu, then click Manage at the top.
3. In the list of models you've downloaded or imported, delete the ones you no longer need.

A model you downloaded comes back when you choose it again in the model menu. A model you imported has to be imported again, from its original file.

Can you delete the files in Finder instead?

Draw Things' documentation doesn't describe removing files by hand, and the shared parts make it easy to break another model. Use Manage for single models.

If you're done with Draw Things altogether, delete its models in Manage before you remove the app. Dragging Draw Things to the Trash leaves its container, models included, on your Mac.

Generating without downloading

Draw Things also offers Cloud Compute, which runs generation on Draw Things' servers. Its documentation suggests it to save space, by using models on the server instead of downloading them. Your prompts and images then leave your Mac, so its documentation recommends local generation for the most privacy.

How big image and video models are

Here's what some common image and video models take, with every file they need to run:

• Stable Diffusion 1.5, ComfyUI2.1 GB
• SDXL Base 8-bit, Draw Things3.9 GB
• SDXL Base, ComfyUI6.9 GB
• SDXL Base, Draw Things7 GB
• FLUX.1 schnell, Draw Things16.6 GB
• Z Image Turbo, Draw Things Exact20.6 GB
• FLUX.1 schnell, ComfyUI full precision34.2 GB
• Wan 2.2 14B video, Draw Things36.8 GB
• Qwen Image 2512, Draw Things Exact55.1 GB

Draw Things sizes from the file sizes it publishes for its downloads; ComfyUI sizes from the Comfy-Org, comfyanonymous and Black Forest Labs files on Hugging Face, October 2026. Models of one family share parts, so two of them take less than the sum.

Three of the larger ones together pass 100 GB.

Why is FLUX so much bigger than Stable Diffusion?

Newer models have far more parameters, and they read prompts with much larger text encoders. Stable Diffusion 1.5 fits in one 2.1 GB file. FLUX.1's main model alone is 23.8 GB at full precision, and its T5 text encoder adds 9.8 GB more.

Why is the same model smaller in Draw Things?

Draw Things offers 8-bit and smaller versions by default, and quantizes the text encoder too. FLUX.1 [schnell] at full precision in ComfyUI, with the 16-bit T5 encoder, is about 34 GB. Draw Things' default 8-bit version, with a 6-bit encoder, is about 16.6 GB.

ComfyUI can use smaller files too. The 8-bit T5 encoder is 4.9 GB instead of 9.8 GB, which brings the same FLUX setup to about 29 GB.

LoRAs are small, but they add up

A LoRA is a small add-on, not a whole model. In Draw Things' download list, the Turbo LoRAs for Stable Diffusion 3.5 Medium are about 0.15 GB and the Lightning LoRAs for Qwen Image about 0.85 GB. The FLUX.1 depth LoRA that Comfy-Org repackages is 1.24 GB.

Collections of dozens of LoRAs, downloaded to try a style once, can still reach tens of gigabytes. Sort the folder by size and date to find the ones you never went back to.

Optional extras: the SDXL refiner

Some models come with optional companions. SDXL's refiner is a second model for the final steps of an image, and Stability AI's model card says the base model can be used on its own. The refiner is 4.5 GB in Draw Things and 6.1 GB as Stability AI's original file.

If you downloaded it and never used it, it's an easy one to remove.

Video models

Video models are among the largest. Draw Things lists Wan 2.2's 14B text-to-video model as two experts, high-noise and low-noise, at 14.8 GB each. With the 6.9 GB text encoder and the small autoencoder they use, that's about 37 GB.

If you tried a video model once, it's likely the single largest thing in your models folder.

DiffusionBee's models

DiffusionBee is a simpler app for Stable Diffusion that runs everything on your Mac. It keeps its models and the images you make in one hidden folder in your home folder, ~/.diffusionbee.

What's in the .diffusionbee folder?

Folder	What it holds	Comes back?
downloaded_assets	Models DiffusionBee downloaded	Yes, DiffusionBee downloads them again
downloads	The same, in older versions	Yes
imported_models	Models you imported, converted to DiffusionBee's own format	Only if you import the original again
images	The images you generated	No, these are your work

DiffusionBee's documentation says there's no button in the app for deleting images, so the images folder keeps growing until you tidy it in Finder.

Why does DiffusionBee download so much at first?

DiffusionBee's documentation says that on first launch it downloads and installs additional data for image generation. Current versions keep their downloads in downloaded_assets; older ones used downloads. Updating the app replaces only the app: its documentation says your history and models are kept.

Imported models are converted

When you add a model in DiffusionBee's settings, it converts the file into its own format and saves the result in imported_models. The model file you chose stays where it was, so keep it only if you might import it again or use it in another app.

Removing DiffusionBee's models

Quit DiffusionBee first. Then open ~/.diffusionbee with Go › Go to Folder and move downloaded_assets, downloads or imported_models to the Trash. DiffusionBee downloads the models it needs again when it next needs them.

Leave the images folder alone unless you've kept the pictures you want somewhere else.

Uninstalling DiffusionBee

DiffusionBee's documentation says to drag the app to the Trash, then remove the .diffusionbee folder for a complete uninstall. That folder holds your generated images too, so copy any you want to keep before you remove it.

DiffusionBee's latest release on GitHub is version 2.5.3, from August 2024. If you've moved on to another app, its folder may be holding models you no longer use.

ComfyUI's models folder

ComfyUI is a node-based tool for building image and video workflows. Unlike Draw Things and DiffusionBee, it uses model files as they come, mostly .safetensors files, from a models folder sorted by type.

Comfy Desktop: installations and the shared library

Comfy Desktop, ComfyUI's Mac app, manages several ComfyUI installations from one window. Each installation lives in ~/ComfyUI-Installs, with its own Python environment, and Comfy's documentation recommends at least 4.85 GB of space for each.

Models go in the shared library at ~/ComfyUI-Shared, which every installation uses by default, along with shared input and output folders. Its app data is in ~/Library/Application Support/Comfy Desktop and its logs in ~/Library/Logs/Comfy Desktop.

The folders inside models

Folder	What goes there
checkpoints	All-in-one model files, such as SD 1.5 and SDXL checkpoints
diffusion_models	Main model files on their own, such as FLUX, Qwen Image and Wan; older guides say unet
text_encoders	Text encoders such as T5 and CLIP; older guides say clip
vae	Autoencoders that turn results into pixels
loras	LoRA add-ons
controlnet	Models that guide an image with edges, depth or poses
upscale_models	Models that enlarge images

Like Draw Things, ComfyUI lets many models share a text encoder or VAE. A file in text_encoders may be needed by every FLUX model you have, so check your workflows before you remove one.

How to remove a model in ComfyUI

Comfy's documentation says there's no uninstall button for models: you remove a model by deleting its file from the folder where you put it. Quit Comfy Desktop, or ComfyUI if you run it on its own, then move the file to the Trash in Finder.

A workflow that used the model can't run until the file is back. Download it again into the same folder to bring it back.

Installations, copies and Forget

Comfy Desktop's installation menu has three actions that matter for space:

• Copy Installation makes a duplicate with its own Python environment and custom nodes, several more gigabytes.
• Forget removes an installation from the list without deleting its files, so the space stays used.
• Uninstall deletes the installation and all its files. Models in the shared library aren't affected.

Each installation's About tab shows how much disk space it uses, so you can compare them before you uninstall one.

Legacy Desktop and installs from GitHub

The earlier ComfyUI Desktop app, which Comfy now calls Legacy Desktop, kept models in a folder you chose during setup. That folder is recorded as basePath in ~/Library/Application Support/ComfyUI/config.json.

When Comfy Desktop migrates a Legacy Desktop installation, it leaves the old one untouched as a backup, copies your inputs and outputs, and creates a new Python environment. Your models stay where they were and become one of the shared folders, so don't remove the old folder without moving them first.

The Legacy Desktop app's own files are separate, in ~/Library/Application Support/ComfyUI, with an extra_models_config.yaml file that lists where it looked for models. Check that list before you remove any folder it names.

If you installed ComfyUI from GitHub, the models folder is inside the folder you installed it in. An extra_model_paths.yaml file can point ComfyUI at more model folders, such as another app's, so check it before deciding where a model lives.

Custom nodes with models of their own

Many newer models run in ComfyUI through community custom nodes, and Comfy's documentation warns that their paths can differ from the usual models folders. Each node's README says where its models go.

GGUF versions of image models, which are smaller, need a custom node too: Comfy's documentation says ComfyUI doesn't read GGUF files by itself. When you remove a custom node, check its README for models it downloaded, because they may stay behind.

Outputs and inputs are yours

ComfyUI's output folder holds every image and video you generated, and input holds the images you fed in. They're your work, not models. Move them somewhere safe before you remove an installation or a whole ComfyUI folder.

Other places image models hide

Image generation isn't limited to these three apps. Python tools and other interfaces keep models in places of their own.

The Hugging Face cache

Hugging Face's Diffusers library, and tools built on it, download models into the shared Hugging Face cache at ~/.cache/huggingface/hub. mflux, an MLX version of FLUX and other image models, also downloads its models from Hugging Face the first time you run one.

The Hugging Face cache on a Mac explains how to see what's there and remove one model at a time.

Stable Diffusion WebUI folders

If you once used AUTOMATIC1111's Stable Diffusion WebUI or a similar interface, its models are inside its own folder, in models/Stable-diffusion, models/Lora and models/VAE. They're easy to forget after moving to ComfyUI or Draw Things.

ComfyUI can read them through extra_model_paths.yaml, so you don't need a second copy. Draw Things and DiffusionBee can't share files: they make their own converted copies.

Apple's Image Playground

Image Playground, Apple's own image app, is part of Apple Intelligence, whose models belong to macOS. macOS keeps them in folders it protects and offers no way to delete them, so they never show up in these apps' folders. How much space Apple Intelligence uses covers what they take and what turning Apple Intelligence off does.

How the three apps compare

	Draw Things	DiffusionBee	ComfyUI
Format on disk	Its own .ckpt files	Its own format	Files as downloaded
Converts what you import	Yes, the original stays	Yes, the original stays	No, it uses the file itself
Shares parts between models	Yes, text encoders and autoencoders	Not documented	Yes, when workflows load the same files
Removes models in the app	Yes, in Manage	Not described in its documentation	No; remove files in Finder

Why the same model can be on your Mac three times

Import one Stable Diffusion checkpoint into Draw Things and DiffusionBee and also use it in ComfyUI, and you have the original, Draw Things' converted copy and DiffusionBee's converted copy: three copies of roughly the same size.

Pick one app for each model where you can. Language models work the same way, as MLX and llama.cpp models on a Mac shows.

How to see which models you have

No single window lists every image model on a Mac. Each app lists its own, and Finder shows the files behind them.

In Draw Things

The Manage window, under Settings and the Model menu, lists every model you've downloaded or imported. It's the place to start, because it shows models by name rather than by file.

In Finder

Open the Draw Things models folder, .diffusionbee or ComfyUI's models folder with Go › Go to Folder, and choose View › as List. Click the Size column to put the largest files first; for folders, turn on Calculate all sizes in View › Show View Options.

The largest files are almost always main models and big text encoders. File names follow the model's name, so the list usually tells you what each one is.

In Comfy Desktop

Each installation's About tab shows its location on disk and how much space it uses. In Desktop Settings, the Storage tab lists every shared model folder, and the folder icon next to each one opens it in Finder.

What removing a model costs

A model you downloaded comes back with another download. Models you made or brought in yourself are another matter.

Downloads come back

Draw Things downloads a model again when you choose it, DiffusionBee downloads its models when it starts, and ComfyUI's model files can be downloaded again from where you got them. On a connection that downloads 100 MB a second, a 17 GB FLUX model takes about three minutes.

Some models need you to sign in and accept a license before you can download them, such as FLUX.1 on Hugging Face. Make sure you still have that access before you remove one.

What doesn't come back

• Models you imported, if you've since removed the original file.
• LoRAs you trained yourself, in Draw Things or any other tool.
• Models their author has taken down. A file shared on a community site can disappear.
• Your images and workflows. They're kept apart from the models, so removing a model never removes them, but removing a whole app folder can.

Back up what's yours first

Before a big tidy-up, copy the things no download can replace to another folder or drive:

• LoRAs you trained, which Draw Things keeps in its models folder with everything else.
• Original files of models you imported, if you might want them again.
• ComfyUI's user folder, which holds your saved workflows, and its output folder.
• DiffusionBee's images folder.

Will removing models speed things up?

No. Models sit on disk until an app loads one. Removing them gives the space back, and nothing more: generation isn't faster, and your Mac isn't either.

Choosing what to keep

Most image model collections grow by trying things. A few decisions keep them in check.

A quick way to decide

If	Then
You have one model in two apps	Keep it in the app you actually use
You have the Exact and 8-bit versions of a model	Keep the 8-bit one unless you can see the difference
You imported a model and it works	Remove the original file, or keep only the original and remove the import
You tried a video model once	It's probably your largest download; remove it
You moved from DiffusionBee to another app	Save the images you want, then remove .diffusionbee
A ComfyUI installation sits unused	Uninstall it in Comfy Desktop; the shared models stay

A worked example

Say Draw Things holds FLUX.1 [schnell] and [dev], SDXL Base, Wan 2.2's 14B video model and the Exact version of Qwen Image 2512, all in their default versions. Because FLUX and SDXL share helper files, the five come to about 128 GB, not the 132 GB their sizes add up to.

If you generate stills with FLUX and SDXL and tried the other two once, deleting Wan and Qwen Image in Manage frees about 92 GB. The three you use stay, in about 36 GB.

Should Time Machine back up your models?

Downloaded models come back with a download, so backing them up mostly fills your backup disk. To leave a folder out, choose Apple menu › System Settings, click General, then Time Machine, click Options, and drag the folder into the list. Click Done.

Leave out only folders of downloads. Draw Things' models folder can also hold LoRAs you trained and models you imported, so copy those somewhere that's backed up before you exclude it. The steps are the same in macOS 26 and macOS 27.

Mistakes to avoid

• Removing shared text encoders or autoencoders by hand. Other models may need them; use Draw Things' Manage window.
• Removing a whole app folder with your images in it. DiffusionBee's and ComfyUI's folders hold your work too.
• Using Forget to free space in Comfy Desktop. It leaves every file in place.
• Removing Legacy Desktop's folder after migrating. Comfy Desktop still reads models from it.
• Removing anything while the app is open. Quit Draw Things or DiffusionBee, and stop ComfyUI, first.
• Pasting cleanup commands from a web page. Nothing here needs Terminal.

Keeping it from growing back

Choose one precision per model, and remove the ones you compared once you've decided. Keep the original files of models you import in one folder, so you know which ones you could bring back, and check your models folders with Get Info every month or two.

Keeping models on another drive

Comfy Desktop handles this best. In its Desktop Settings, the Storage tab lists shared model folders: add one on an external drive with Add Shared Directory, and choose Make primary to save new models there.

Draw Things' documentation doesn't describe moving its models folder, so leave the container where it is. Where local AI tools keep their models in general is in the guide to local AI models.

How the app does it

Wombit explains these folders like everything else on your disk. Select one in the map, the app's home screen, to see what it is, how safe it is to remove and what removing it costs, from the catalog's entries.

• ~/Library/Containers/com.liuliu.draw-things/Data/Documents/Models

  Draw Things models

  Image and video generation models downloaded or imported in Draw Things.

  Review first · Explained only: the app shows you how

• ~/.diffusionbee/{downloaded_assets,downloads,imported_models}

  DiffusionBee models

  Stable Diffusion models that DiffusionBee downloaded or that you imported into it.

  Review first · Moves it to the Trash

• ~/.cache/huggingface/hub

  Hugging Face cache

  Models and datasets downloaded from Hugging Face, shared by every library and tool that uses them.

  Review first · Moves it to the Trash

All three are Review first: models are yours to choose, and imported models may not come back, so nothing starts checked and none of it counts in the headline number, which counts only Safe items.

DiffusionBee's model folders and the Hugging Face cache move to the Trash after the review sheet, where you can put them back, and DiffusionBee's images folder is never part of them. Draw Things' models are explained with the steps to remove them in Draw Things, which knows which files each model shares.

ComfyUI's folders have no catalog entry. Items no entry explains can get a Best guess written on your Mac by Apple Intelligence, labelled as a guess, and they're never offered for removal.

Seeing what your models take is free. Moving them to the Trash is PRO, one $4.99 payment with lifetime access. Every Pro tool's review is free to open, so you see what would go and why before you buy.